行为分析技术研究及其应用
文献类型：学位论文
| 作者 | 董秋雷
|
| 学位类别 | 工学博士 |
| 答辩日期 | 2008-05-23 |
| 授予单位 | 中国科学院研究生院 |
| 授予地点 | 中国科学院自动化研究所 |
| 导师 | 胡占义 ; 吴毅红 |
| 关键词 | 行为分析 视频分割 异常行为检测 行为识别 定位与高度测量 子空间学习 Behavior analysis video segmentation abnormal behavior detection behavior recognition localization and height measurement subspace learning |
| 其他题名 | A Study on Behavior Analysis and its Applications |
| 学位专业 | 模式识别与智能系统 |
| 中文摘要 | 人的运动分析是计算机视觉领域的一个重要研究方向，具有重要的理论价值和应用价值。人运动分析的最终目标是自动分析和理解人的个人行为、人与人之间及人与其它目标之间的交互行为，因此行为分析是运动分析研究中最具挑战的主题。本文针对行为分析技术中存在的一些问题进行了探索和研究，主要工作归纳如下：1.提出了一种基于贪心相似度量的视频分割方法。该方法首先计算每个动作序列中运动前景的时空主曲线，然后对这些主曲线分别建立隐马尔可夫模型，最后通过最大化贪心相似度量将对应多个动作的长主曲线分割成若干个对应于单个行为的曲线段，并同时实现了动作识别。2.提出了一种基于二次曲线的行为识别方法。该方法基于改进的贪心交换法则跟踪视频序列中运动前景的特征点，然后将跟踪得到的点集拟合成二次曲线，最后提取二次曲线的6个不变量特征作为行为模板，实现基本动作的识别。实验结果表明该方法识别精度较高且计算量低。3.提出了一种基于损失函数的逐点运动图表述，并结合AdaBoost法则，进一步提出了一种基于逐点运动图的异常行为检测和行为识别方法。这种逐点运动图表述本身是一幅HSV颜色空间中的彩色图像，逐点运动图中每个像素的3个颜色分量H、S、V分别表示这个点的运动方向、速度和周期。因此通过图像和视频处理技术可以很容易地从这种运动表述中提取有较高区分能力的特征。实验结果证明了逐点运动图表述在异常行为检测和行为识别中的有效性。4.提出了一种基于视频的实时自动人体高度测量方法。该方法首先在视频序列中的每帧图像上提取一种新的头部特征点以及一种新的脚部特征点。然后根据这些特征点建立约束方程求出近似的人体高度，并同时在视频序列中跟踪双脚。最后基于获得的双脚跟踪结果引入一条关于特征点所对应空间点的几何约束进一步优化测量结果，从而实现实时自动人体定位和高度测量。该方法有效地利用了视频序列中包含的运动信息，有较强的鲁棒性和较高的测量精度，既能有效地处理透视镜头下的视频又能处理鱼眼镜头下的视频，而且计算量很低，可以实现实时测量。5.提出了一种有监督的子空间学习方法——局部线性保持映射。该方法通过局部线性拟合近似地学习出数据集的非线型流形结构。它不仅保持邻域内异类点之间的距离不变以及全局类中心点之间的距离不变，同时还尽量减小邻域内同类点之间的距离。因此，该方法可以描述原始数据空间中的一些非线型属性，并能有效地分类新的测试数据。 |
| 英文摘要 | Human motion analysis is an important research direction in the domain of computer vision. Since the goal of human motion analysis is to analyze and understand human behaviors automatically, behavior analysis is the most challenging research topic in human motion analysis. This thesis is focused on some key aspects of behavior analysis, and the main contributions are summarized as: 1. A novel segmentation method based on greedy similarity measure is proposed to segment long spatial-temporal video sequences automatically. Firstly, a principal curve of motion region along frames of a video sequence is constructed to represent trajectory. Then from the constructed principal curves of trajectories of predefined gestures, HMMs are applied to their modeling. For a long input video sequence, the greedy similarity measure is adapted to automatically segment it into gestures along with gesture recognition. 2. A novel method based on quadratic curves is presented for human gesture recognition. In this method, the feature points of the motion foreground are kept tracked by a modified Greedy Exchange algorithm respectively. Then these points are fitted into a quadratic curve and 6 invariants from this quadratic curve are computed. Lastly, the gesture models are learnt from the invariant feature of gesture samples and an input gesture is recognized by comparing its feature vector with those of gesture models. 3. A motion representation called Pointwise Motion Image (PMI) is proposed, which is under the form of a color image in the HSV color space, where the color components of each pixel represent the pointwise motion speed, pointwise motion orientation and pointwise motion duration respectively. And then a method combining the PMI and AdaBoost is presented for behavior recognition and abnormality detection. 4. A real-time automatic method is proposed for measuring the height of human body. Firstly, a new kind of head feature point and a new kind of feet feature point are extracted in each frame of the input video sequence. Then, all these extracted feature points are used to construct several constraint equations for computing the approximate human height, and simultaneously the feet are tracked in the video sequence. Finally, based on the tracking results, the measurement result is refined with an additional geometric constraint on the spatial points corresponding to the extracted feature points along with human location. 5. A supervised subspace learning method called Locally Linear Preserving Mapping (LLPM) is proposed. LLPM learns the global structure of nonlinear manifold approximately from locally linear fits. It can not only preserve local distances of neighboring points with different labels and the distances of global class centers, but it can also reduce the distances of neighboring points in the same class. |
| 语种 | 中文 |
| 其他标识符 | 200518014628046 |
| 源URL | [http://ir.ia.ac.cn/handle/173211/6066]  |
| 专题 | 毕业生_博士学位论文 |
| 推荐引用方式 GB/T 7714 | 董秋雷. 行为分析技术研究及其应用[D]. 中国科学院自动化研究所. 中国科学院研究生院. 2008. |
入库方式： OAI收割
来源：自动化研究所
浏览0
下载0
收藏0
其他版本
除非特别说明，本系统中所有内容都受版权保护，并保留所有权利。